Kernels:

kernels-community
/

rwkv

Trusted publisher

kernels-bot commited on May 4

Commit

eaf0f88

verified ·

1 Parent(s): ffc2322

Uploaded using `kernel-builder`.

Browse files

Files changed (34) hide show

.gitattributes +9 -0
build/torch210-cxx11-cu126-aarch64-linux/_ops.py +3 -3
build/torch210-cxx11-cu126-aarch64-linux/{_rwkv_cuda_86f75d9.abi3.so → _rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch210-cxx11-cu126-aarch64-linux/metadata.json +1 -1
build/torch210-cxx11-cu128-aarch64-linux/_ops.py +3 -3
build/{torch211-cxx11-cu128-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch210-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch210-cxx11-cu128-aarch64-linux/metadata.json +1 -1
build/torch210-cxx11-cu130-aarch64-linux/_ops.py +3 -3
build/{torch211-cxx11-cu130-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch210-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch210-cxx11-cu130-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu126-aarch64-linux/_ops.py +3 -3
build/torch211-cxx11-cu126-aarch64-linux/{_rwkv_cuda_86f75d9.abi3.so → _rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch211-cxx11-cu126-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu128-aarch64-linux/_ops.py +3 -3
build/{torch210-cxx11-cu128-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch211-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch211-cxx11-cu128-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu130-aarch64-linux/_ops.py +3 -3
build/{torch210-cxx11-cu130-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch211-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} +1 -1
build/torch211-cxx11-cu130-aarch64-linux/metadata.json +1 -1
build/torch212-cxx11-cu126-aarch64-linux/__init__.py +170 -0
build/torch212-cxx11-cu126-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu126-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so +3 -0
build/torch212-cxx11-cu126-aarch64-linux/metadata.json +15 -0
build/torch212-cxx11-cu126-aarch64-linux/rwkv/__init__.py +26 -0
build/torch212-cxx11-cu130-aarch64-linux/__init__.py +170 -0
build/torch212-cxx11-cu130-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so +3 -0
build/torch212-cxx11-cu130-aarch64-linux/metadata.json +17 -0
build/torch212-cxx11-cu130-aarch64-linux/rwkv/__init__.py +26 -0
build/torch212-cxx11-cu132-aarch64-linux/__init__.py +170 -0
build/torch212-cxx11-cu132-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu132-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so +3 -0
build/torch212-cxx11-cu132-aarch64-linux/metadata.json +17 -0
build/torch212-cxx11-cu132-aarch64-linux/rwkv/__init__.py +26 -0

.gitattributes CHANGED Viewed

@@ -159,3 +159,12 @@ build/torch211-cxx11-cu130-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs di
 build/torch212-cxx11-cu126-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu132-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text

 build/torch212-cxx11-cu126-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu132-x86_64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu126-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu126-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu126-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu132-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-cxx11-cu126-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/{_rwkv_cuda_86f75d9.abi3.so → _rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:db4d37ddbafb2f617859098993d32d43db0fc9db93e98505efa79b4bd282a030
 size 2232072

 version https://git-lfs.github.com/spec/v1
+oid sha256:a0a87c0d1d20d0268d41ce17fc2956188dc741d477c88c3a3b097663b974079a
 size 2232072

build/torch210-cxx11-cu126-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch210-cxx11-cu128-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/{torch211-cxx11-cu128-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch210-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2bb5666854e7d30a79b3665bdb4dc1b414e1e99e0f9b6c6b830bb5e040c4723e
 size 2428792

 version https://git-lfs.github.com/spec/v1
+oid sha256:57c5971cfc8211061308e458b61b1cca798fe240f9d9b764207e45962b9f9649
 size 2428792

build/torch210-cxx11-cu128-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch210-cxx11-cu130-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/{torch211-cxx11-cu130-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch210-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:35da2e270b6779a634d2e41b07f385952e3689b8d5aa66da14f70c029388b3f3
 size 2433400

 version https://git-lfs.github.com/spec/v1
+oid sha256:63f96928ddcb5057297d6fa3fb938752f76630ee104589a2e0c79634ac332769
 size 2433400

build/torch210-cxx11-cu130-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu126-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/torch211-cxx11-cu126-aarch64-linux/{_rwkv_cuda_86f75d9.abi3.so → _rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:da8acbdf94af9a4245064830b237dbb79a5ed61f66bf5baec9753257c0269d60
 size 2232072

 version https://git-lfs.github.com/spec/v1
+oid sha256:d91f9c8ef6e83975d97417e1e63e78fd04009ca2ce6a4bae673e4814585d9fc0
 size 2232072

build/torch211-cxx11-cu126-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu128-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/{torch210-cxx11-cu128-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch211-cxx11-cu128-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6e06058a91fcd84154ad74740166945559d8d1a02b3cadef47149d0bf5dc673a
 size 2428792

 version https://git-lfs.github.com/spec/v1
+oid sha256:65811a3820a7cbb2846cc83c4e670739c90433c44fa8a7346b69de0c3469591f
 size 2428792

build/torch211-cxx11-cu128-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu130-aarch64-linux/_ops.py CHANGED Viewed

@@ -1,9 +1,9 @@
 import torch
-from . import _rwkv_cuda_86f75d9
-ops = torch.ops._rwkv_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rwkv_cuda_86f75d9::{op_name}"

 import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/{torch210-cxx11-cu130-aarch64-linux/_rwkv_cuda_86f75d9.abi3.so → torch211-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:be9f74b4a43cafcd2c41f7424dc65f63825654fc0ad077f8b3d5ce0ba146eede
 size 2433400

 version https://git-lfs.github.com/spec/v1
+oid sha256:f3b59a8b28fe59c6778e9c3f46c0854ed4b770fbc57fa2a2e9648ff66de0db30
 size 2433400

build/torch211-cxx11-cu130-aarch64-linux/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "rwkv",
-  "id": "_rwkv_cuda_86f75d9",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

 {
   "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch212-cxx11-cu126-aarch64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,170 @@

+from ._ops import ops
+from typing import Tuple, Any
+# Use a broad Tensor alias to avoid importing torch at import time.
+from torch import Tensor
+def forward(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (float32).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward(w, u, k, v, y)
+def forward_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (bfloat16 inputs/outputs, float32 ``w``).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward_bf16(w, u, k, v, y)
+def forward_with_state(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (float32).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state(w, u, k, v, y, s)
+def forward_with_state_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (bfloat16 inputs/outputs, float32 ``w`` and ``s``).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state_bf16(w, u, k, v, y, s)
+def backward(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (float32).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.float32`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward(w, u, k, v, y, gy, gw, gu, gk, gv)
+def backward_bf16(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (bfloat16 inputs/outputs/gradients, float32 ``w``).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.bfloat16`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward_bf16(w, u, k, v, y, gy, gw, gu, gk, gv)
+def _validate_device_match(tensors: Tuple[Tensor, ...]) -> None:
+    """Minimal runtime validation that all tensors live on the same CUDA device."""
+    if not tensors:
+        return
+    device = tensors[0].device
+    if not device.type == "cuda":
+        raise RuntimeError("RWKV CUDA ops require CUDA tensors")
+    for t in tensors[1:]:
+        if t.device != device:
+            raise RuntimeError("All tensors must be on the same CUDA device")
+__all__ = [
+    "forward",
+    "forward_bf16",
+    "forward_with_state",
+    "forward_with_state_bf16",
+    "backward",
+    "backward_bf16",
+]

build/torch212-cxx11-cu126-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/torch212-cxx11-cu126-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:02c639b66eb5ef2f4384c0680d86f1fe160b1c9a8963d521cba10083b250a07a
+size 2232120

build/torch212-cxx11-cu126-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,15 @@

+{
+  "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch212-cxx11-cu126-aarch64-linux/rwkv/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu130-aarch64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,170 @@

+from ._ops import ops
+from typing import Tuple, Any
+# Use a broad Tensor alias to avoid importing torch at import time.
+from torch import Tensor
+def forward(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (float32).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward(w, u, k, v, y)
+def forward_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (bfloat16 inputs/outputs, float32 ``w``).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward_bf16(w, u, k, v, y)
+def forward_with_state(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (float32).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state(w, u, k, v, y, s)
+def forward_with_state_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (bfloat16 inputs/outputs, float32 ``w`` and ``s``).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state_bf16(w, u, k, v, y, s)
+def backward(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (float32).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.float32`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward(w, u, k, v, y, gy, gw, gu, gk, gv)
+def backward_bf16(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (bfloat16 inputs/outputs/gradients, float32 ``w``).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.bfloat16`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward_bf16(w, u, k, v, y, gy, gw, gu, gk, gv)
+def _validate_device_match(tensors: Tuple[Tensor, ...]) -> None:
+    """Minimal runtime validation that all tensors live on the same CUDA device."""
+    if not tensors:
+        return
+    device = tensors[0].device
+    if not device.type == "cuda":
+        raise RuntimeError("RWKV CUDA ops require CUDA tensors")
+    for t in tensors[1:]:
+        if t.device != device:
+            raise RuntimeError("All tensors must be on the same CUDA device")
+__all__ = [
+    "forward",
+    "forward_bf16",
+    "forward_with_state",
+    "forward_with_state_bf16",
+    "backward",
+    "backward_bf16",
+]

build/torch212-cxx11-cu130-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/torch212-cxx11-cu130-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:b1fd38632d1a9c1b1ec491840e0f3b6862be3b376441b69404a4cc1c01d4321c
+size 2433408

build/torch212-cxx11-cu130-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,17 @@

+{
+  "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch212-cxx11-cu130-aarch64-linux/rwkv/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu132-aarch64-linux/__init__.py ADDED Viewed

	@@ -0,0 +1,170 @@

+from ._ops import ops
+from typing import Tuple, Any
+# Use a broad Tensor alias to avoid importing torch at import time.
+from torch import Tensor
+def forward(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (float32).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward(w, u, k, v, y)
+def forward_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor) -> None:
+    """RWKV WKV forward pass (bfloat16 inputs/outputs, float32 ``w``).
+    Runs the CUDA kernel and writes the result into ``y`` in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y))
+    ops.forward_bf16(w, u, k, v, y)
+def forward_with_state(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (float32).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.float32``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state(w, u, k, v, y, s)
+def forward_with_state_bf16(w: Tensor, u: Tensor, k: Tensor, v: Tensor, y: Tensor, s: Tensor) -> None:
+    """RWKV WKV forward pass with persistent state (bfloat16 inputs/outputs, float32 ``w`` and ``s``).
+    Runs the CUDA kernel using and updating state ``s`` and writes the result into ``y``.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u: Input tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        k: Key tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        v: Value tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        y: Output tensor, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+        s: Stateful tensor, shape ``[B, C]``, dtype ``torch.float32`` (updated in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B`` and ``C``; ``y`` shares ``[B, T, C]`` with inputs.
+    """
+    _validate_device_match((w, u, k, v, y, s))
+    ops.forward_with_state_bf16(w, u, k, v, y, s)
+def backward(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (float32).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.float32``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.float32`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.float32`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward(w, u, k, v, y, gy, gw, gu, gk, gv)
+def backward_bf16(
+    w: Tensor,
+    u: Tensor,
+    k: Tensor,
+    v: Tensor,
+    y: Tensor,
+    gy: Tensor,
+    gw: Tensor,
+    gu: Tensor,
+    gk: Tensor,
+    gv: Tensor,
+) -> None:
+    """RWKV WKV backward pass (bfloat16 inputs/outputs/gradients, float32 ``w``).
+    Writes gradients into the provided tensors in-place.
+    Args:
+        w: Decay weights, shape ``[C]``, dtype ``torch.float32``.
+        u, k, v, y: Forward-pass tensors, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gy: Gradient of ``y``, shape ``[B, T, C]``, dtype ``torch.bfloat16``.
+        gw: Gradient for ``w``, shape ``[C]``, dtype ``torch.bfloat16`` (written in-place).
+        gu, gk, gv: Gradients for ``u``, ``k``, ``v`` respectively, shape ``[B, T, C]``, dtype ``torch.bfloat16`` (written in-place).
+    Notes:
+        - All tensors must be on the same CUDA device.
+        - Shapes must agree on ``B``, ``T`` and ``C``.
+    """
+    _validate_device_match((w, u, k, v, y, gy, gw, gu, gk, gv))
+    ops.backward_bf16(w, u, k, v, y, gy, gw, gu, gk, gv)
+def _validate_device_match(tensors: Tuple[Tensor, ...]) -> None:
+    """Minimal runtime validation that all tensors live on the same CUDA device."""
+    if not tensors:
+        return
+    device = tensors[0].device
+    if not device.type == "cuda":
+        raise RuntimeError("RWKV CUDA ops require CUDA tensors")
+    for t in tensors[1:]:
+        if t.device != device:
+            raise RuntimeError("All tensors must be on the same CUDA device")
+__all__ = [
+    "forward",
+    "forward_bf16",
+    "forward_with_state",
+    "forward_with_state_bf16",
+    "backward",
+    "backward_bf16",
+]

build/torch212-cxx11-cu132-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _rwkv_cuda_dfdeb22
+ops = torch.ops._rwkv_cuda_dfdeb22
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_rwkv_cuda_dfdeb22::{op_name}"

build/torch212-cxx11-cu132-aarch64-linux/_rwkv_cuda_dfdeb22.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:16a4fe19a41f0d5b06690c2cca4965a35cd78f6503c5bda15ded23929b70bb8a
+size 2501256

build/torch212-cxx11-cu132-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,17 @@

+{
+  "name": "rwkv",
+  "id": "_rwkv_cuda_dfdeb22",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch212-cxx11-cu132-aarch64-linux/rwkv/__init__.py ADDED Viewed

	@@ -0,0 +1,26 @@

+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))